Cross-scene power load prediction method, device and system under large model, and medium

Abstract

The embodiment of the invention provides a cross-scene power load prediction method, device and system under a large model, and a medium, and relates to the technical field of power load prediction. The method comprises the following steps: generating a cross-scene power load association feature set according to specific power consumption type power load data of a source scene and a target scene and corresponding power consumption influence factor data; according to the cross-scene power load association feature set, adjusting a weight parameter of a pre-trained power load prediction large model, and converting a prediction parameter of a source scene into a load prediction adaptation parameter for adapting to a target scene through transfer learning; and performing real-time load feature optimization and prediction calculation on the target scene by using the large model after the weight parameter adjustment and the load prediction adaptive parameter, and generating a prediction result of the power load of the specific power consumption type of the target scene. According to the invention, the accuracy and adaptability of cross-scene power load prediction are improved.

Description

Technical Field

[0001] This application relates to the field of power load forecasting technology, specifically to a method, device, system, and storage medium for cross-scenario power load forecasting under a large model. Background Technology

[0002] In power systems, accurate power load forecasting is crucial for power dispatch, resource allocation, and ensuring the stable operation of the power system. Traditional power load forecasting methods mainly rely on the analysis and modeling of historical load data under a single scenario. For example, they use time series analysis methods, such as the autoregressive moving average (ARMA) model, to predict future power load based solely on the power load values ​​of a certain region over a past period.

[0003] However, the aforementioned single-scenario prediction methods have significant limitations. Electricity load in different scenarios is influenced by a variety of factors, which manifest differently in different scenarios. For example, electricity load in industrial scenarios is mainly affected by factors such as industrial production shifts and equipment operating status; while electricity load in residential scenarios is closely related to residents' daily routines and weather conditions. Single-scenario prediction methods cannot fully consider these differences, resulting in low prediction accuracy across different scenarios.

[0004] Furthermore, when applying a mature prediction model from one scenario to another, traditional methods often cannot be directly transferred due to the differences between scenarios. Extensive data collection and model retraining are required, which is not only time-consuming and resource-intensive but also difficult to adapt quickly to the needs of different scenarios. Therefore, existing power load forecasting methods have significant shortcomings in cross-scenario applications and cannot meet the increasingly complex and diverse needs of the power system. Summary of the Invention

[0005] This application provides a method, device, system, and storage medium for cross-scenario power load prediction under a large model.

[0006] The first aspect of this application provides a method for cross-scenario power load forecasting under a large model, including:

[0007] Based on the specific electricity load data of the source and target scenarios and the corresponding electricity influencing factors data, a cross-scenario electricity load association feature set is generated.

[0008] Based on the cross-scenario power load association feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning.

[0009] By utilizing the large model with adjusted weight parameters and load forecasting adaptation parameters, real-time load characteristic optimization and forecasting calculations are performed for the target scenario, generating power load forecasting results for specific electricity consumption types in the target scenario.

[0010] In an optional embodiment of this application, a cross-scenario power load correlation feature set is generated based on power load data of specific power consumption types in the source and target scenarios and corresponding power consumption influencing factor data, including:

[0011] Extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the source scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the time period association feature of power load in the source scenario.

[0012] Extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the target scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the time period association feature of power load in the target scenario;

[0013] The power load time period association features of the source scenario and the power load time period association features of the target scenario are mapped one-to-one according to the time period identifier. The similarity parameter between the power load time period association features of the source scenario and the power load time period association features of the target scenario under the corresponding time period identifier is calculated to generate cross-scenario load time period association features.

[0014] In an optional embodiment of this application, load value sequences corresponding to the same time period identifier are extracted from the power load data set of a specific power consumption type in the source scenario, the variation range of load values ​​under the same time period identifier on different dates is calculated, and the time period association feature of the power load in the source scenario is generated, including:

[0015] From the power load data set of specific power consumption type in the source scenario, all load data with the same time period identifier are selected to form the load value sequence of the source scenario. Based on the load values ​​of two adjacent dates in the load value sequence of the source scenario, the load change amplitude of all adjacent dates is calculated to form the load change amplitude sequence.

[0016] Based on the load change magnitude sequence, we obtain the upward trend characteristic value, downward trend characteristic value and stable trend characteristic value, and integrate them to form the time period correlation characteristics of the power load in the source scenario.

[0017] In an optional embodiment of this application, a cross-scenario power load correlation feature set is generated based on power load data of specific power consumption types in the source and target scenarios and corresponding power consumption influencing factor data, including:

[0018] Extract the correlation between meteorological data and industrial production shift data from the data of electricity consumption influencing factors in the source scenario, calculate the load change rate corresponding to industrial production shifts under different meteorological data, and generate the correlation characteristics of electricity consumption influencing factors in the source scenario.

[0019] Extract the correlation between meteorological data and residents' daily routine data from the data on electricity consumption influencing factors in the target scenario, calculate the load change rate corresponding to residents' daily routine under different meteorological data, and generate the correlation features of electricity consumption influencing factors in the target scenario.

[0020] The correlation features of power consumption influencing factors in the source scenario and the correlation features of power consumption in the target scenario are mapped one-to-one according to meteorological data types. Complementary parameters of the correlation features of power consumption influencing factors in the source scenario and the correlation features of power consumption in the target scenario under the corresponding meteorological data type are calculated to generate cross-scenario load influencing factor correlation features.

[0021] In an optional embodiment of this application, the weight parameters of a pre-trained large-scale power load prediction model are adjusted based on a cross-scenario power load association feature set, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including:

[0022] The cross-scenario load time period correlation features are input into the large model. The large model performs hierarchical analysis on the source scenario time period features, target scenario time period features, and similarity parameters in the cross-scenario load time period correlation features, and identifies significant time period feature dimensions that affect the accuracy of load prediction.

[0023] Calculate the contribution ratio of the significant time period feature dimension in the source scene and the target scene, and adjust the feature extraction weight of the corresponding dimension in the large model according to the contribution ratio.

[0024] In an optional embodiment of this application, the weight parameters of a pre-trained large-scale power load prediction model are adjusted based on a cross-scenario power load association feature set, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including:

[0025] The cross-scenario load influencing factor correlation features are input into the large model. The large model performs hierarchical analysis on the source scenario influencing factor features, target scenario influencing factor features and complementary parameters in the cross-scenario load influencing factor correlation features, and identifies the significant influencing factor feature dimensions that affect the accuracy of load prediction.

[0026] Calculate the complementarity coefficients of the feature dimensions of significant influencing factors in the source and target scenarios, and adjust the feature extraction weights of the corresponding dimensions in the large model based on the complementarity coefficients.

[0027] In an optional embodiment of this application, the weight parameters of a pre-trained large-scale power load prediction model are adjusted based on a cross-scenario power load association feature set, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including:

[0028] Based on the time period characteristics and influencing factor characteristics of the source and target scenarios after weighting, and based on the scenario difference parameters in the cross-scenario power load association feature set, the fusion ratio of time period characteristics and influencing factor characteristics is adjusted to generate cross-scenario fused load characteristics.

[0029] Input the cross-scenario fused load features into the large model to verify the load prediction error of the cross-scenario fused load features. If the error exceeds the preset range, readjust the feature extraction weights of the corresponding dimensions until the prediction error meets the preset range.

[0030] In an optional embodiment of this application, based on the time-period characteristics and influencing factor characteristics of the source and target scenarios after weighting, and based on the scenario difference parameters in the cross-scenario power load association feature set, the fusion ratio of the time-period characteristics and influencing factor characteristics is adjusted to generate cross-scenario fused load characteristics, including:

[0031] The time-period features and influencing factor features of the output are obtained from the large model after weight adjustment, and a cross-scenario fused load feature vector is generated based on the time-period features and influencing factor features;

[0032] The cross-scenario fusion load feature vectors corresponding to all time period identifiers and electricity consumption influencing factor types are integrated to form cross-scenario fusion load features.

[0033] In an optional embodiment of this application, the weight parameters of a pre-trained large-scale power load prediction model are adjusted based on a cross-scenario power load association feature set, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including:

[0034] The source scenario power load prediction parameters are extracted from the large model. The time period feature extraction weights in the source scenario power load prediction parameters are subjected to scenario difference adaptation processing. Based on the load peak difference in the scenario difference features, the adjustment ratio of the time period feature extraction weights is calculated. If the load peak of the target scenario in a certain time period is higher than that of the corresponding time period in the source scenario, the time period feature extraction weights of the corresponding time period are increased, and vice versa. The adapted time period feature extraction weights are obtained.

[0035] In an optional embodiment of this application, the weight parameters of a pre-trained large-scale power load prediction model are adjusted based on a cross-scenario power load association feature set, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including:

[0036] The feature extraction weights of influencing factors in the power load forecasting parameters of the source scenario are subjected to scenario difference adaptation processing. Based on the difference in load change rate in the scenario difference features, the adjustment coefficient of the feature extraction weights of influencing factors is calculated. If the load change rate under a certain meteorological data in the target scenario is higher than the change rate under the corresponding meteorological data in the source scenario, the feature extraction weight of the influencing factor corresponding to the meteorological data is increased, and vice versa, the feature extraction weight of the influencing factor corresponding to the meteorological data is decreased, so as to obtain the adapted feature extraction weights of influencing factors.

[0037] In an optional embodiment of this application, the weight parameters of a pre-trained large-scale power load prediction model are adjusted based on a cross-scenario power load association feature set, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including:

[0038] The feature fusion coefficients in the power load prediction parameters of the source scenario are subjected to scenario difference adaptation processing. Based on the difference in load fluctuation cycle in the scenario difference features, the correction value of the feature fusion coefficient is calculated. If the load fluctuation cycle of the target scenario is longer than that of the source scenario, the proportion of time period features in the fusion is increased and the feature fusion coefficient is corrected. Conversely, the proportion of influencing factor features in the fusion is increased and the feature fusion coefficient is corrected to obtain the adapted feature fusion coefficient.

[0039] The prediction calculation parameters in the source scenario power load prediction parameters are subjected to scenario difference adaptation processing. Based on the comprehensive impact of load peak difference, load change rate difference, and load fluctuation cycle difference in the scenario difference characteristics, the compensation value of the prediction calculation parameters is calculated. The compensation value is superimposed on the source scenario prediction calculation parameters to obtain the adapted prediction calculation parameters.

[0040] The adapted time period feature extraction weights, influencing factor feature extraction weights, feature fusion coefficients and prediction calculation parameters are integrated to form load prediction adaptation parameters for specific electricity consumption types in the target scenario.

[0041] The load forecasting adaptation parameters are substituted into the historical load data samples of the specific electricity consumption type in the target scenario, and the forecasting verification calculation is performed. If the deviation between the forecasting result and the historical actual load data exceeds the preset threshold, the parameter adjustment ratio, adjustment coefficient, correction value and compensation value are recalculated until the deviation meets the preset threshold.

[0042] In an optional embodiment of this application, using a large model with adjusted weight parameters and load forecasting adaptation parameters, real-time load characteristic optimization and forecasting calculations are performed for the target scenario to generate power load forecasting results for specific electricity consumption types in the target scenario, including:

[0043] Acquire real-time power load data for specific power consumption types in the target scenario and corresponding real-time power consumption influencing factor data. Obtain real-time time period correlation features and real-time influencing factor correlation features through a large model, and further generate real-time fused load features.

[0044] Based on real-time fusion load characteristics, load numerical prediction is performed through a large model to form a power load prediction sequence for specific electricity consumption types in the target scenario.

[0045] The peak load, valley load, and average load for each forecast period are extracted from the power load forecast sequence for specific power consumption types in the target scenario and used as forecast feature parameters. The power load forecast sequence for specific power consumption types in the target scenario is then integrated with the forecast feature parameters to generate the power load forecast result for specific power consumption types in the target scenario.

[0046] In an optional embodiment of this application, using a large model with adjusted weight parameters and load forecasting adaptation parameters, real-time load characteristic optimization and forecasting calculations are performed for the target scenario to generate power load forecasting results for specific electricity consumption types in the target scenario, including:

[0047] Based on the power load forecast results for specific power consumption types in the target scenario, load adjustment instructions are generated for power dispatching in the target scenario. The content of the load adjustment instructions corresponds to the load time period variation pattern in the power load forecast results for specific power consumption types in the target scenario.

[0048] A second aspect of this application provides a cross-scenario power load forecasting device under a large model, including a processor and a memory storing program instructions. The processor is configured to execute the cross-scenario power load forecasting method under a large model as described in the first aspect of this application when running the program instructions.

[0049] A third aspect of this application provides a system comprising:

[0050] The system itself; and,

[0051] As in the second aspect of the embodiments of this application, the cross-scenario power load prediction device under the large model is installed on the system body.

[0052] A fourth aspect of the embodiments of this application provides a computer-readable storage medium storing program instructions that, when executed, cause a computer to perform a cross-scenario power load forecasting method under a large model as described in the first aspect of the embodiments of this application.

[0053] The cross-scenario power load prediction method, apparatus, system, and storage medium under the large model provided in the embodiments of this application have the following beneficial effects:

[0054] This application embodiment collects power load data of specific power consumption types and corresponding power consumption influencing factors from source and target scenarios, and generates a cross-scenario power load correlation feature set. It considers the time-period correlation features and power consumption influencing factor correlation features of power load under different scenarios. Then, it performs hierarchical interactive processing of power load features on the pre-trained power load prediction model, adjusting the extraction weights related to power load time-period features and power consumption influencing factor features in the model. This allows the model to better adapt to the feature extraction needs of different scenarios. Through transfer learning, it performs transfer learning power scenario parameter adaptation processing, converting the prediction parameters of the source scenario into load prediction adaptation parameters adapted to the target scenario, combining scenario difference features. This achieves rapid model transfer and adaptation between different scenarios, avoiding the high cost and long time consumption of retraining the model. By inputting real-time data from the target scenario into the weighted model and combining it with the adaptation parameters for real-time load feature optimization and prediction calculation, accurate power load prediction results for the target scenario can be generated, along with load adjustment instructions for power dispatch in the target scenario. This effectively improves the operating efficiency and stability of the power system and adapts to the complex and ever-changing actual needs of the power system. Attached Figure Description

[0055] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0056] Figure 1 This is a schematic diagram of the cross-scenario power load prediction method under a large model provided in the embodiments of this application;

[0057] Figure 2 This is a schematic diagram of a cross-scenario power load prediction device under a large model provided in the embodiments of this application.

[0058] Figure label:

[0059] 800: Cross-scenario power load prediction device under large model; 801: Processor; 802: Memory; 803: Communication interface; 804: Bus. Detailed Implementation

[0060] To make the technical solutions and advantages of the embodiments of this application clearer, the exemplary embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other.

[0061] Figure 1 This is a schematic diagram of a cross-scenario power load prediction method under a large model provided in this application embodiment. The method can be executed in the system or in a server or terminal device that is connected to the system.

[0062] Combination Figure 1 As shown in the embodiments of this application, a cross-scenario power load forecasting method under a large model is provided, including:

[0063] S100 generates a cross-scenario power load association feature set based on the power load data of specific power consumption types in the source scenario and the target scenario, as well as the corresponding power consumption influencing factor data.

[0064] In an optional embodiment of this application, load value sequences corresponding to the same time period identifier are extracted from the power load data set of a specific power consumption type in the source scenario, and the variation range of load values ​​under the same time period identifier on different dates is calculated to generate power load time period association features of the source scenario; load value sequences corresponding to the same time period identifier are extracted from the power load data set of a specific power consumption type in the target scenario, and the variation range of load values ​​under the same time period identifier on different dates is calculated to generate power load time period association features of the target scenario; the power load time period association features of the source scenario and the power load time period association features of the target scenario are mapped one-to-one according to the time period identifier, and the similarity parameter between the power load time period association features of the source scenario and the power load time period association features of the target scenario under the corresponding time period identifier is calculated to generate cross-scenario load time period association features.

[0065] In an optional embodiment of this application, all load data with the same time period identifier are selected from the power load data set of specific power consumption type in the source scenario to form a load value sequence of the source scenario. Based on the load values ​​of two adjacent dates in the load value sequence of the source scenario, the load change amplitude of all adjacent dates is calculated to form a load change amplitude sequence. Based on the load change amplitude sequence, the upward trend feature value, the downward trend feature value, and the stable trend feature value are obtained and integrated to form the time period correlation feature of the power load in the source scenario.

[0066] In an optional embodiment of this application, the correlation between meteorological data and industrial production shift data is extracted from the source scenario electricity consumption influencing factor data, and the load change rate corresponding to industrial production shifts under different meteorological data is calculated to generate source scenario electricity consumption influencing factor correlation features; the correlation between meteorological data and residential daily life data is extracted from the target scenario electricity consumption influencing factor data, and the load change rate corresponding to residential daily life under different meteorological data is calculated to generate target scenario electricity consumption influencing factor correlation features; the source scenario electricity consumption influencing factor correlation features and the target scenario electricity consumption influencing factor correlation features are mapped one-to-one according to meteorological data type, and the complementary parameters of the source scenario electricity consumption influencing factor correlation features and the target scenario electricity consumption influencing factor correlation features under the corresponding meteorological data type are calculated to generate cross-scenario load influencing factor correlation features.

[0067] Specifically, in this embodiment, a power load data set of a specific power consumption type in the source scenario and a power load data set of a specific power consumption type in the target scenario are collected. Simultaneously, power consumption influencing factor data corresponding to the power load data set of the specific power consumption type in the source scenario and the power load data set of the specific power consumption type in the target scenario are collected. Based on the time period correlation characteristics and power consumption influencing factor correlation characteristics of the power load in the power load data set of the specific power consumption type in the source scenario and the power load data set of the specific power consumption type in the target scenario, a cross-scenario power load correlation feature set is generated.

[0068] In this embodiment, the above process is specifically illustrated using an industrial power consumption scenario as the source scenario and a residential power consumption scenario as the target scenario. Industrial power consumption scenarios and residential power consumption scenarios are two typical power consumption scenarios, and their power consumption patterns and influencing factors are significantly different. Through cross-scenario power load prediction and analysis, power dispatch optimization under different scenarios can be better achieved.

[0069] In this embodiment of the application, regarding step S100:

[0070] S110, collect a set of power load data for a specific power consumption type in the source scenario through a power data acquisition terminal. The set of power load data for a specific power consumption type in the source scenario includes load values ​​for the specific power consumption type in the source scenario at different dates and time periods. Each load value corresponds to a time period identifier for the recorded collection time point.

[0071] Specifically, for the source scenario of industrial electricity consumption, the power data acquisition terminal can be a smart meter or load monitoring device installed in the factory's power distribution system. These terminal devices collect industrial electricity load data according to a preset sampling frequency (e.g., once every 15 minutes). The collected data includes load values ​​for different dates (e.g., weekdays, weekends, holidays) and different time periods (e.g., 8:00-9:00 AM, 2:00-3:00 PM). To accurately distinguish different time periods, each load value is recorded with a corresponding time period identifier. These time period identifiers can be divided according to the time intervals of a day; for example, a 24-hour day can be divided into 96 15-minute time periods, each assigned a unique identifier number. For example, time period 1 corresponds to 00:00-00:15, time period 2 corresponds to 00:15-00:30, and so on. Through these identifiers, the load data for different time periods can be categorized and analyzed.

[0072] S120, collect a set of power load data for a specific power consumption type in the target scenario through a power data acquisition terminal of the same type. The set of power load data for a specific power consumption type in the target scenario includes load values ​​for the specific power consumption type in the target scenario at different dates and time periods. Each load value also corresponds to a time period identifier for the recorded collection time point.

[0073] The target scenario is residential electricity consumption. The same type of power data acquisition terminal as the source scenario is used, such as installing the same model acquisition device on the main distribution cabinet in the residential community or on the smart meters in users' homes, to ensure data consistency and accuracy. The collected residential electricity load data also includes load values ​​for different dates (weekdays, weekends, holidays, etc.) and different time periods. Each load value also corresponds to a time period identifier recorded at the time of collection. The division of this time period identifier is completely consistent with the source scenario; for example, it is also divided into 96 15-minute time periods, each corresponding to the same identifier number. In this way, the load data of the source and target scenarios are comparable in terms of time period division, facilitating subsequent cross-scenario time period correlation feature analysis.

[0074] S130, synchronously collect electricity consumption influencing factor data corresponding to the power load data set of specific electricity consumption type in the source scene through environmental data acquisition equipment. The electricity consumption influencing factor data includes meteorological data and industrial production shift data at the time of collection in the source scene.

[0075] In the source scenario of industrial power consumption, environmental data acquisition equipment includes weather stations installed within the factory premises and interfaces connected to the factory's production management system. Weather stations collect electricity-related meteorological data such as temperature, humidity, air pressure, wind speed, and solar radiation intensity. This meteorological data affects air conditioning load, ventilation load, and the operating efficiency of certain production equipment during industrial production, thus influencing the overall power load. Industrial production shift data is obtained through the connection with the factory's production management system, including the start-up shifts of each production line (e.g., morning, afternoon, and evening shifts), the working hours of each shift, and production plans (e.g., whether production is at full capacity, whether equipment maintenance is required). This production shift data directly reflects the activity level of industrial production and has a significant impact on the magnitude and trend of power load. The collected meteorological data and industrial production shift data are synchronized with the corresponding power load data in time; that is, the load data at each point in time has corresponding meteorological and production shift data.

[0076] S140, synchronously collect electricity consumption influencing factor data corresponding to the electricity load data set of a specific electricity consumption type in the target scenario through environmental data acquisition equipment of the same type. The electricity consumption influencing factor data includes meteorological data and residents' daily life data at the time of collection in the target scenario.

[0077] For the target scenario of residential electricity consumption, environmental data acquisition equipment of the same type as that used in the source scenario is used. Meteorological data is collected through weather stations installed in residential communities, and the collected meteorological parameters, such as temperature, humidity, and wind speed, are consistent with those in the source scenario to facilitate effective comparison during cross-scenario analysis. Data on residents' daily routines is collected through various methods, such as connecting with the property management system of the residential community to obtain information on occupancy rates and population flow; statistical analysis of residents' electricity consumption behavior to obtain typical patterns in residential electricity consumption periods, such as morning wake-up time, midday cooking time, and evening lighting and entertainment time; and combining factors such as holiday schedules and seasonal changes to comprehensively form residential routine data. All of the above data also corresponds synchronously with residential electricity load data in time.

[0078] S150: Extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the source scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the power load time period association feature of the source scenario; extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the target scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the power load time period association feature of the target scenario.

[0079] Furthermore, regarding step S150:

[0080] S151, extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the source scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the time period association feature of power load in the source scenario.

[0081] Furthermore, regarding step S151:

[0082] S151-1: Select all load data with the same time period identifier from the power load data set of specific power consumption type in the source scenario, arrange them in chronological order of collection date, and form the source scenario load value sequence corresponding to the time period identifier.

[0083] In the source scenario of industrial electricity consumption, taking a time period identified as "8:00-9:00 AM on weekdays" (assuming this time period is identified as time period 33) as an example, all load data for time period 33 are filtered from the power load data set of specific electricity consumption types in the source scenario. These data come from different weekdays. Then, the above load data are arranged in chronological order of collection dates (e.g., from Monday to Friday, then to the following Monday, etc.) to form the source scenario load value sequence corresponding to this time period identifier. For example, this source scenario load value sequence may contain industrial electricity load data for all weekdays from 8:00 AM to 9:00 AM over the past year.

[0084] S151-2: Extract the load values ​​of two adjacent dates from the source scenario load value sequence, calculate the difference between the load value of the later date and the load value of the earlier date, and obtain the load difference between adjacent dates.

[0085] For the source scenario load value sequence formed above, the load values ​​of adjacent dates are extracted sequentially. For example, the load value of the nth date and the load value of the (n+1)th date in the sequence are calculated by subtracting the load value of the nth date from the load value of the (n+1)th date. This difference is the load difference between adjacent dates. By calculating the load difference between adjacent dates, a preliminary understanding of the load changes during that period can be obtained.

[0086] S151-3, calculate the ratio of the load difference between adjacent dates to the load value of the previous date to obtain the load change range between adjacent dates.

[0087] Divide the load difference between adjacent dates obtained in step S151-2 by the load value of the previous date to obtain the load variation range between adjacent dates. This variation range reflects the proportion of load change relative to the previous date, and can more intuitively show the degree of load fluctuation. For example, if the load value of the previous date is L1, and the load difference between adjacent dates is ΔL, then the load variation range between adjacent dates is ΔL / L1.

[0088] S151-4. Following the steps above, calculate the load change magnitude of all adjacent dates in the source scenario load value sequence in turn to form the load change magnitude sequence corresponding to the source scenario for that time period.

[0089] Following the methods in steps S151-2 and S151-3, all adjacent date combinations in the source scenario load value sequence are calculated to obtain a series of adjacent date load change amplitudes. Arranging these change amplitudes in chronological order of dates forms the load change amplitude sequence corresponding to the source scenario for that time period.

[0090] S151-5, perform trend analysis on the load change magnitude sequence corresponding to the source scenario for this time period, identify the rising, falling, and stable intervals of the change magnitude, and record the start and end dates of each interval. Further:

[0091] S151-5(1): Extract all load change amplitude values ​​from the load change amplitude sequence corresponding to the source scene for that time period, arrange them in chronological order of the collection date, and form an amplitude sequence list.

[0092] Extract all values ​​from the load change magnitude sequence corresponding to the time period identifier of the source scenario. These values ​​are already arranged in chronological order according to the collection date, thus directly forming a magnitude sequence list.

[0093] S151-5(2): Set an upward determination threshold and a stable determination threshold, wherein the upward determination threshold is a positive number and the stable determination threshold is a value range close to zero.

[0094] Based on the general fluctuations in industrial electricity load, thresholds for determining an upward trend and a stable trend are set. The upward trend threshold is a positive number; when the load change exceeds this threshold, the load is considered to be on an upward trend. The stable trend threshold is a value close to zero; when the load change falls within this range, the load is considered to be in a stable state. When the load change is less than the negative upward trend threshold, the load is considered to be on a downward trend. The setting of these thresholds needs to be adjusted based on historical data and experience to adapt to the characteristics of different industrial scenarios.

[0095] S151-5(3): Starting from the first value in the amplitude sequence list, determine the attribute of each value in turn: if the value is greater than the rising threshold, it is marked as rising amplitude value; if the value is within the range of the stable threshold, it is marked as stable amplitude value; if the value is less than the negative rising threshold, it is marked as falling amplitude value.

[0096] Starting with the first value in the amplitude sequence list, each value is evaluated sequentially. Each value is compared with the rising threshold and the stabilizing threshold, and is marked as a rising amplitude value, a stabilizing amplitude value, or a falling amplitude value based on the comparison result. For example, if a value is 0.05, the rising threshold is 0.03, and the stabilizing threshold is -0.02 to 0.02, then this value is greater than the rising threshold and is marked as a rising amplitude value.

[0097] S151-5(4): Starting from the first value, find the starting and ending positions of consecutively increasing values. The date corresponding to the starting position is the starting date of the rising interval, and the date corresponding to the ending position is the ending date of the rising interval. Record the starting and ending dates of the rising interval.

[0098] Starting from the first value in the amplitude sequence list, search for a sequence of consecutive increasing amplitude values. When the first increasing amplitude value is encountered, record its corresponding date as the start date of the increasing interval; continue searching until a non-increasing amplitude value (i.e., a stable amplitude value or a decreasing amplitude value) is encountered. The date corresponding to the previous increasing amplitude value at this point is the end date of the increasing interval. Record this pair of start and end dates as one increasing interval.

[0099] S151-5(5): Continue to find the starting and ending positions of consecutive decreasing values. The date corresponding to the starting position is the starting date of the decreasing interval, and the date corresponding to the ending position is the ending date of the decreasing interval. Record the starting and ending dates of the decreasing interval.

[0100] After completing the search for the rising interval, continue searching the amplitude sequence list for a sequence of consecutive falling amplitude values. The search method is similar to that for the rising interval. When the first falling amplitude value is encountered, record its corresponding date as the start date of the falling interval; continue searching until a non-falling amplitude value is encountered. The date corresponding to the previous falling amplitude value is the end date of the falling interval, and this interval is recorded.

[0101] S151-5(6): Next, find the starting and ending positions of the consecutively occurring stable amplitude values. The date corresponding to the starting position is the starting date of the stable interval, and the date corresponding to the ending position is the ending date of the stable interval. Record the starting and ending dates of the stable interval.

[0102] After finding the decreasing interval, continue searching for sequences of consecutive stable amplitude values. Similarly, record the start and end dates as the stable intervals.

[0103] S151-5(7): If there are adjacent amplitude values ​​of different attributes that alternate in the amplitude sequence list, then the single stable amplitude value in the middle is assigned to the adjacent rising interval; if there are multiple alternating amplitude values, then the corresponding segment sequence is marked as a transition interval and is not classified for the time being.

[0104] In the amplitude sequence list, adjacent amplitude values ​​of different attributes may alternate, for example, an increasing amplitude value followed by a stable amplitude value, and then another increasing amplitude value. In such cases, if the intermediate stable amplitude value is a single value, it is classified into the adjacent increasing interval. If multiple amplitude values ​​of different attributes alternate (such as increasing, stable, decreasing, stable, increasing, etc.), this sequence is marked as a transition interval and is not temporarily classified into an increasing, decreasing, or stable interval.

[0105] S151-5(8): Perform a second judgment on the amplitude value of the transition interval, calculate the average value of all amplitude values ​​in the transition interval. If the average value is greater than the rising judgment threshold, the transition interval is assigned to the adjacent rising interval; if the average value is within the stability judgment threshold range, the transition interval is assigned to the adjacent stable interval; if the average value is less than the negative rising judgment threshold, the transition interval is assigned to the adjacent falling interval.

[0106] For a sequence labeled as a transition interval, calculate the average of all amplitude values ​​within that interval. Compare this average to the rising threshold and the stationary threshold. If the average is greater than the rising threshold, the transition interval is assigned to an adjacent rising interval; if the average is within the stationary threshold range, it is assigned to an adjacent stationary interval; if the average is less than a negative rising threshold, it is assigned to an adjacent falling interval. This two-step process allows for a more accurate classification of transition intervals.

[0107] S151-5(9): After completing the interval classification of all amplitude values, organize the start and end dates of all rising intervals, falling intervals and stable intervals to form a list of load change amplitude intervals corresponding to the source scenario time period identifier, which is used to generate the source scenario power load time period sub-features corresponding to the time period identifier.

[0108] The start and end dates of all the rising, falling, and stable intervals identified in the above steps are compiled to form a complete list of load change ranges. This list of load change ranges shows the trend of load change amplitude across different date intervals under the specified time period in the source scenario.

[0109] Continuing with step S151:

[0110] S151-6 Calculate the average value of all load changes within the rising interval as the characteristic value of the rising trend for that period; calculate the average value of all load changes within the falling interval as the characteristic value of the falling trend for that period; calculate the average value of the absolute values ​​of all load changes within the stable interval as the characteristic value of the stable trend for that period.

[0111] For each increasing interval in the list of load change ranges, calculate the average of all load change ranges within that interval. Then, average these averages again (if there are multiple increasing intervals) to obtain the upward trend characteristic value for that time period. Similarly, for each decreasing interval, calculate and average the average of the load change ranges within that interval to obtain the downward trend characteristic value. For stable intervals, calculate and average the average of the absolute values ​​of the load change ranges within that interval to obtain the stable trend characteristic value. These characteristic values ​​can summarize the trend characteristics of load changes under the specified time period.

[0112] S151-7 integrates the load change amplitude sequence, upward trend characteristic value, downward trend characteristic value and stable trend characteristic value of the time period to form the source scenario power load time period sub-feature corresponding to the time period identifier.

[0113] The load change amplitude sequence obtained in step S151-4, along with the upward trend characteristic value, downward trend characteristic value, and stable trend characteristic value calculated in step S11516, are integrated to form the sub-feature of the power load period corresponding to the source scenario for that period. This sub-feature contains detailed information on the load change process and overall trend of that period.

[0114] S151-8, following the same steps, process all time period identifiers in the power load data set of specific power consumption type in the source scenario in sequence, and generate the source scenario power load time period sub-feature corresponding to each time period identifier.

[0115] For all time period identifiers (such as 96 time period identifiers) in the power load data set of a specific power consumption type in the source scenario, they are processed according to the method of steps S151-1 to S151-7 to generate the source scenario power load time period sub-feature corresponding to each time period identifier.

[0116] S151-9, arrange the source scenario power load time period sub-features corresponding to all time period identifiers according to the time order of the time period identifiers to form the source scenario power load time period association features. The source scenario power load time period association features include the change amplitude sequence and trend feature value of each time period.

[0117] All source scenario power load time period sub-features corresponding to all time period identifiers are arranged according to the time order of the time period identifiers (e.g., from time period 1 to time period 96), and combined to form the source scenario power load time period association feature. This source scenario power load time period association feature integrates the load change information of all time periods in the source scenario.

[0118] Continuing with step S150:

[0119] S152, extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the target scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the time period association feature of power load in the target scenario.

[0120] The target scenario is a residential electricity consumption scenario, and the process of generating the time-period correlation features of the target scenario's electricity load is similar to that of the source scenario. Taking the time period identified as "7 PM to 8 PM on weekdays" (assuming this time period is identified as time period 70) as an example, all load data for this time period are filtered out from the residential electricity load data set, arranged by date to form a load value sequence, the load difference and change amplitude between adjacent dates are calculated to form a change amplitude sequence, trend analysis is performed to obtain rising, falling, and stable intervals and trend characteristic values, and these are integrated to form the target scenario's electricity load time period sub-features for this time period. Finally, all the time period sub-features are arranged in chronological order to form the target scenario's electricity load time period correlation features.

[0121] Continuing with step S100:

[0122] S160, the power load time period association features of the source scenario and the power load time period association features of the target scenario are mapped one-to-one according to the time period identifier. The similarity parameter between the power load time period association features of the source scenario and the power load time period association features of the target scenario under the corresponding time period identifier is calculated to generate cross-scenario load time period association features.

[0123] The sub-features with the same time period identifier in the power load time period association features of the source scenario and the target scenario are mapped one-to-one. For each corresponding time period identifier, a similarity parameter is calculated. The similarity parameter can be calculated using various methods, such as calculating the correlation coefficient between the change magnitude sequences of the two time period sub-features; the closer the correlation coefficient is to 1, the more similar the change trends are. The degree of difference between upward trend feature values, downward trend feature values, and stable trend feature values ​​is also calculated; the smaller the difference, the higher the similarity. These similarity indicators are combined to form the similarity parameter under the corresponding time period identifier. The similarity parameters corresponding to all time period identifiers are then integrated with the time period sub-features of the source and target scenarios to generate cross-scenario load time period association features.

[0124] S170: Extract the correlation between meteorological data and industrial production shift data from the source scenario electricity consumption influencing factor data, calculate the load change rate corresponding to industrial production shifts under different meteorological data, and generate the correlation features of source scenario electricity consumption influencing factors; extract the correlation between meteorological data and residential daily life data from the target scenario electricity consumption influencing factor data, calculate the load change rate corresponding to residential daily life under different meteorological data, and generate the correlation features of target scenario electricity consumption influencing factors.

[0125] Furthermore, regarding step S170:

[0126] S171, extract the correlation between meteorological data and industrial production shift data from the source scenario electricity consumption influencing factor data, calculate the load change rate corresponding to industrial production shifts under different meteorological data, and generate the correlation characteristics of source scenario electricity consumption influencing factors.

[0127] In the source scenario of industrial electricity consumption, both meteorological data (such as temperature) and industrial production shift data (such as whether it is a normal production shift or an overtime shift) affect the power load. First, the electricity consumption influencing factor data of the source scenario is analyzed to extract the correlation between meteorological data and industrial production shift data. For example, in hot weather, industrial production may require more cooling equipment to be turned on, and the load will be higher if it is a normal production shift; while in cold weather and during non-production shifts, the load may be lower. Then, for different meteorological data types (such as high temperature, low temperature, suitable temperature, etc.), the load change rate under different industrial production shifts is calculated. The load change rate can be calculated by comparing the difference between the average load of different production shifts under the same meteorological data and the average load of a benchmark shift (such as a normal production shift). Integrating the load change rates corresponding to industrial production shifts under different meteorological data generates the correlation characteristics of the electricity consumption influencing factors in the source scenario.

[0128] S172, extract the correlation between meteorological data and residents' daily routine data from the data on electricity consumption influencing factors in the target scenario, calculate the load change rate corresponding to residents' daily routine under different meteorological data, and generate the correlation features of electricity consumption influencing factors in the target scenario.

[0129] For the target scenario of residential electricity consumption, the correlation between meteorological data and residents' daily routine data is also analyzed. For example, during hot summer evenings (peak electricity consumption periods in residents' daily routines), residents use cooling devices such as air conditioners, leading to increased load; while during the pleasant temperatures of spring and autumn and off-peak electricity consumption periods, the load is relatively low. For different types of meteorological data, the load variation rate is calculated for different periods of residents' daily routines (e.g., morning, noon, evening). The load variation rate is calculated by comparing the average load for different periods under the same meteorological data with the average load for a baseline period (e.g., off-peak electricity consumption periods). These load variation rates are integrated to generate the correlation characteristics of electricity consumption influencing factors in the target scenario.

[0130] Continuing with step S100:

[0131] S180: The correlation features of power consumption influencing factors in the source scenario and the correlation features of power consumption influencing factors in the target scenario are mapped one-to-one according to meteorological data types. The complementary parameters of the correlation features of power consumption influencing factors in the source scenario and the correlation features of power consumption influencing factors in the target scenario under the corresponding meteorological data type are calculated to generate cross-scenario load influencing factor correlation features.

[0132] The system maps the common meteorological data types (e.g., high temperature, low temperature) in the correlation characteristics of electricity consumption influencing factors in the source and target scenarios. For each meteorological data type, a complementary parameter is calculated. This complementary parameter measures the degree of complementarity between the electricity consumption influencing factors in the source and target scenarios under that meteorological data type. For example, if, under high-temperature meteorological data, the load change rate in the source scenario is mainly affected by industrial production shifts, while the load change rate in the target scenario is mainly affected by residents' daily routines, and their influence trends are opposite or complementary, then the complementary parameter can reflect this relationship. The method for calculating the complementary parameter can be to analyze the differences and synergies in the load change rates of the two scenarios and quantify them through comprehensive evaluation indicators. The complementary parameters corresponding to all meteorological data types are then integrated with the correlation characteristics of electricity consumption influencing factors in the source and target scenarios to generate cross-scenario load influencing factor correlation characteristics.

[0133] S190, the cross-scenario load time period correlation features and the cross-scenario load influencing factor correlation features are integrated to form a cross-scenario power load correlation feature set. Each feature in the cross-scenario power load correlation feature set includes correlation parameters between the source scenario and the target scenario, as well as scenario difference parameters.

[0134] The cross-scenario load time-period correlation features include load correlation information and similarity parameters (correlation parameters) between the source and target scenarios at different time periods, while also reflecting the differences in their time-period characteristics (scenario difference parameters). The cross-scenario load influencing factor correlation features include correlation information and complementary parameters (correlation parameters) of electricity influencing factors under different meteorological data types, as well as scenario difference parameters. Integrating these two correlation features forms the cross-scenario power load correlation feature set. Each feature in this cross-scenario power load correlation feature set comprehensively reflects the correlation and differences in power load between the source and target scenarios.

[0135] See also Figure 1 :

[0136] S200 adjusts the weight parameters of the pre-trained power load prediction model based on the cross-scenario power load association feature set, and transforms the prediction parameters of the source scenario into load prediction adaptation parameters for the target scenario through transfer learning.

[0137] In an optional embodiment of this application, cross-scenario load time period correlation features are input into a large model. The large model performs hierarchical analysis on the source scenario time period features, target scenario time period features, and similarity parameters in the cross-scenario load time period correlation features to identify significant time period feature dimensions that affect load prediction accuracy. The contribution ratio of the significant time period feature dimensions in the source scenario and the target scenario is calculated, and the feature extraction weight of the corresponding dimension in the large model is adjusted according to the contribution ratio.

[0138] In an optional embodiment of this application, the cross-scenario load influencing factor correlation features are input into a large model. The large model performs hierarchical analysis on the source scenario influencing factor features, target scenario influencing factor features, and complementary parameters in the cross-scenario load influencing factor correlation features to identify the significant influencing factor feature dimensions for load prediction accuracy. The complementary coefficients of the significant influencing factor feature dimensions in the source scenario and the target scenario are calculated, and the feature extraction weights of the corresponding dimensions in the large model are adjusted according to the complementary coefficients.

[0139] In an optional embodiment of this application, based on the time period characteristics and influencing factor characteristics of the source and target scenarios after weight adjustment, and based on the scenario difference parameters in the cross-scenario power load association feature set, the fusion ratio of the time period characteristics and influencing factor characteristics is adjusted to generate cross-scenario fused load characteristics; the cross-scenario fused load characteristics are input into a large model to verify the load prediction error of the cross-scenario fused load characteristics; if the error exceeds a preset range, the feature extraction weights of the corresponding dimensions are readjusted until the prediction error meets the preset range.

[0140] In an optional embodiment of this application, the time period features and influencing factor features of the output are obtained from the large model after the weights are adjusted, and a cross-scenario fused load feature vector is generated based on the time period features and influencing factor features; the cross-scenario fused load feature vectors corresponding to all time period identifiers and electricity consumption influencing factor types are integrated to form cross-scenario fused load features.

[0141] In an optional embodiment of this application, source scenario power load prediction parameters are extracted from a large model. Scenario difference adaptation processing is performed on the time period feature extraction weights in the source scenario power load prediction parameters. Based on the load peak difference in the scenario difference features, the adjustment ratio of the time period feature extraction weights is calculated. If the load peak of a certain time period in the target scenario is higher than the corresponding time period in the source scenario, the time period feature extraction weight corresponding to that time period is increased; otherwise, the time period feature extraction weight corresponding to that time period is decreased, thus obtaining the adapted time period feature extraction weights.

[0142] In an optional embodiment of this application, scenario difference adaptation processing is performed on the feature extraction weights of influencing factors in the power load prediction parameters of the source scenario. Based on the difference in load change rate in the scenario difference features, the adjustment coefficient of the feature extraction weights of influencing factors is calculated. If the load change rate under a certain meteorological data in the target scenario is higher than the change rate under the corresponding meteorological data in the source scenario, the feature extraction weights of the influencing factors corresponding to that meteorological data are increased; otherwise, the feature extraction weights of the influencing factors corresponding to that meteorological data are decreased, thus obtaining the adapted feature extraction weights of influencing factors.

[0143] In an optional embodiment of this application, scenario difference adaptation processing is performed on the feature fusion coefficients in the power load prediction parameters of the source scenario. Based on the load fluctuation cycle difference in the scenario difference features, a correction value for the feature fusion coefficients is calculated. If the load fluctuation cycle of the target scenario is longer than that of the source scenario, the proportion of time period features in the fusion is increased, and the feature fusion coefficient is corrected; conversely, the proportion of influencing factor features in the fusion is increased, and the feature fusion coefficient is corrected, thus obtaining the adapted feature fusion coefficients. Scenario difference adaptation processing is also performed on the prediction calculation parameters in the power load prediction parameters of the source scenario, based on the load peak difference, load change rate difference, and load fluctuation cycle difference in the scenario difference features. The comprehensive impact of period differences is considered, and the compensation value of the prediction calculation parameters is calculated. The compensation value is then superimposed on the prediction calculation parameters of the source scenario to obtain the adapted prediction calculation parameters. The adapted time period feature extraction weights, influencing factor feature extraction weights, feature fusion coefficients, and prediction calculation parameters are integrated to form load prediction adaptation parameters adapted to the specific electricity consumption type of the target scenario. The load prediction adaptation parameters are substituted into the historical load data samples of the specific electricity consumption type of the target scenario, and prediction verification calculation is performed. If the deviation between the prediction result and the historical actual load data exceeds the preset threshold, the parameter adjustment ratio, adjustment coefficient, correction value, and compensation value are recalculated until the deviation meets the preset threshold.

[0144] In this embodiment of the application, a hierarchical interactive processing of power load features of the pre-trained power load prediction model is performed. The cross-scenario power load correlation feature set is input into the pre-trained power load prediction model, and the extraction weights related to power load time period features and electricity consumption influencing factor features in the pre-trained power load prediction model are adjusted to obtain the power load prediction model after weight adjustment.

[0145] Furthermore, regarding step S200:

[0146] S210, Obtain a pre-trained large-scale power load prediction model. The pre-trained large-scale power load prediction model includes a load feature input layer, a time period feature extraction layer, an influencing factor feature extraction layer, a feature fusion layer, and a prediction output layer. The time period feature extraction layer is used to extract the time period variation features of power load, and the influencing factor feature extraction layer is used to extract the load correlation features corresponding to the power consumption influencing factors.

[0147] The pre-trained large-scale power load forecasting model is obtained by training on a large amount of power load data (which may include historical data from multiple scenarios). The load feature input layer receives external input feature data, such as a set of cross-scenario power load correlation features. The time-period feature extraction layer typically consists of multiple neural network layers (such as convolutional layers and recurrent layers), which extracts features related to power load time-period changes from the input feature data, such as the patterns of load peak and valley occurrences at different times, and load trends over time. The influencing factor feature extraction layer also consists of corresponding neural network layers, used to extract the correlation features between power consumption influencing factors (such as meteorological factors, production shifts, and daily routines) and load from the feature data, analyzing how these factors affect the magnitude and variation of the load. The feature fusion layer fuses the features output from the time-period feature extraction layer and the influencing factor feature extraction layer, comprehensively considering the combined effects of time periods and influencing factors on the load. The prediction output layer predicts the power load based on the fused features and outputs the prediction result.

[0148] S220, the cross-scenario load time period association features in the cross-scenario power load association feature set are input into the time period feature extraction layer of the pre-trained power load prediction model. The time period feature extraction layer performs hierarchical analysis on the source scenario time period features, target scenario time period features and similarity parameters in the cross-scenario load time period association features, and identifies the time period feature dimensions that have a significant impact on the load prediction accuracy.

[0149] Cross-scenario load time-period correlation features are input into the time-period feature extraction layer. This layer first performs preliminary preprocessing on the input features, such as normalization, to adapt them to the model's input requirements. Then, through multi-layer neural network computation, it performs hierarchical analysis of the source scenario time-period features, the target scenario time-period features, and the similarity parameters between them. During this analysis, the model automatically learns the degree of influence of different time-period feature dimensions on load prediction accuracy. For example, the load change magnitude sequence for certain time periods may contain more prediction information, while certain trend feature values ​​have a greater impact on the prediction results. By analyzing the outputs and gradient changes of each layer of the model, those time-period feature dimensions that significantly affect load prediction accuracy can be identified.

[0150] S230: Based on the identified significant time period feature dimensions, calculate the contribution ratio of the significant time period feature dimension in the source scene and the target scene, and adjust the feature extraction weight of the corresponding dimension in the time period feature extraction layer according to the contribution ratio, so that the weight allocation matches the contribution of cross-scene time period features.

[0151] For each identified significant time-period feature dimension, its contribution in both the source and target scenes is calculated. The contribution can be measured by the magnitude of the feature's effect on the model's prediction process, for example, by calculating its influence on the model's loss function. Then, the ratio of the source scene's contribution to the target scene's contribution is calculated. If this ratio is greater than 1, it indicates that the source scene's contribution to that time-period feature dimension is greater, and the feature extraction weight for the corresponding dimension in the source scene can be appropriately increased when adjusting the weights; conversely, if the ratio is less than 1, the weight for the corresponding dimension in the target scene should be increased. This method ensures that the weight allocation of the time-period feature extraction layer matches the contribution of cross-scene time-period features, thereby improving the model's efficiency in utilizing cross-scene time-period features.

[0152] S240, the cross-scenario load influencing factor correlation features in the cross-scenario power load correlation feature set are input into the influencing factor feature extraction layer of the pre-trained power load prediction model. The influencing factor feature extraction layer performs hierarchical analysis on the source scenario influencing factor features, target scenario influencing factor features and complementary parameters in the cross-scenario load influencing factor correlation features, and identifies the influencing factor feature dimensions that have a significant impact on the load prediction accuracy.

[0153] Similarly, the cross-scenario load influencing factors are correlated with the features of the influencing factors and are input into the influencing factor feature extraction layer. This layer performs hierarchical analysis on the features of influencing factors from the source scenario, the target scenario, and complementary parameters. Through layer-by-layer processing by the neural network, the impact of different influencing factor feature dimensions (such as load change rates under different meteorological data types) on load prediction accuracy is analyzed. Combined with feedback information from the model training process, the influencing factor feature dimensions that significantly affect prediction accuracy are identified.

[0154] S250: Based on the identified significant influencing factor feature dimensions, calculate the complementarity coefficient of the dimension in the source and target scenarios, and adjust the feature extraction weights of the corresponding dimensions in the influencing factor feature extraction layer according to the complementarity coefficients, so that the weight allocation matches the complementarity of cross-scenario influencing factor features.

[0155] For the feature dimensions of significant influencing factors, their complementarity coefficients in the source and target scenarios are calculated. The complementarity coefficient reflects the degree of complementarity between the source and target scenarios in this dimension. A higher complementarity indicates that the features of both scenarios can complement each other, jointly improving prediction accuracy. The feature extraction weights for the corresponding dimension of the influencing factor feature extraction layer are adjusted based on the magnitude of the complementarity coefficient. If the complementarity coefficient is high, it indicates that both the source and target scenario features in this dimension have significant predictive value and should be given higher weights. If the complementarity coefficient is low, it may mean that one scenario's features are more dominant; therefore, the weight allocation is adjusted accordingly to match the complementarity of the cross-scenario influencing factor features.

[0156] S260, the time period features output by the time period feature extraction layer after weight adjustment and the influencing factor features output by the influencing factor feature extraction layer after weight adjustment are input into the feature fusion layer. The feature fusion layer adjusts the fusion ratio of the time period features and the influencing factor features based on the scenario difference parameters in the cross-scenario power load association feature set, and generates cross-scenario fused load features.

[0157] Furthermore, regarding step S260:

[0158] S261, Obtain the output time period features from the time period feature extraction layer after weight adjustment. The time period features include feature vectors corresponding to each time period identifier, and the dimension of each feature vector is consistent with the feature dimension of the time period feature extraction layer.

[0159] After weight adjustment, the time-period feature extraction layer processes the cross-scene load time-period correlation features and outputs time-period features. The time-period features are represented in the form of feature vectors, with each time-period identifier corresponding to a feature vector. The dimension of the feature vectors is determined by the structure of the time-period feature extraction layer. For example, if the number of neurons in the last layer of the time-period feature extraction layer is m, then the dimension of each feature vector is m.

[0160] S262, obtain the output influencing factor features from the influencing factor feature extraction layer after weight adjustment. The influencing factor features include feature vectors corresponding to each type of electricity consumption influencing factor, and the dimension of each feature vector is consistent with the feature dimension of the influencing factor feature extraction layer.

[0161] Similarly, the feature extraction layer of influencing factors after adjusting the weights outputs the features of influencing factors. Each type of electricity influencing factor (such as meteorological factors, production shift factors, daily routine factors, etc.) corresponds to a feature vector. The dimension of the feature vector is determined by the structure of the feature extraction layer of influencing factors, which is assumed to be n.

[0162] S263, extract scenario difference parameters from the cross-scenario power load association feature set. The scenario difference parameters include the load ratio difference between the source scenario and the target scenario under each time period identifier, and the load contribution difference under each type of electricity consumption influencing factor.

[0163] Scenario difference parameters are an important component of the cross-scenario power load correlation feature set. Load share difference refers to the difference between the proportion of load in the source scenario and the proportion of load in the target scenario in the total daily load under the same time period identifier; load contribution difference refers to the difference between the proportion of the contribution of the same power consumption influencing factor to the load in the source scenario and the proportion of the same factor to the load in the target scenario, under the same type of power consumption influencing factor.

[0164] S264: Calculate the absolute value of the load proportion difference under each time period identifier, sum them and take the average value to obtain the comprehensive value of time period difference; calculate the absolute value of the load contribution difference under each type of electricity consumption influencing factor, sum them and take the average value to obtain the comprehensive value of influencing factor difference.

[0165] For each time period identifier, the absolute value of the load share difference is calculated. Then, the absolute values ​​of all time period identifiers are summed, and the result is divided by the total number of time period identifiers to obtain the comprehensive value of the time period difference. Similarly, for each type of electricity consumption influencing factor, the absolute value of the load contribution difference is calculated, summed, and divided by the total number of electricity consumption influencing factor types to obtain the comprehensive value of the influencing factor difference. These two comprehensive values ​​reflect the overall degree of difference between the source scenario and the target scenario in terms of time period characteristics and influencing factor characteristics, respectively.

[0166] S265, add the comprehensive value of time period differences to the comprehensive value of differences in influencing factors to obtain the total difference value; calculate the ratio of the comprehensive value of time period differences to the total difference value as the fusion weight ratio of time period features; calculate the ratio of the comprehensive value of differences in influencing factors to the total difference value as the fusion weight ratio of influencing factor features.

[0167] The total difference value is the sum of the combined difference value of time periods and the combined difference value of influencing factors. The fusion weight of time period characteristics equals the combined difference value of time periods divided by the total difference value, and the fusion weight of influencing factor characteristics equals the combined difference value of influencing factors divided by the total difference value. Initially, the sum of these two weights is 1.

[0168] S266. If the comprehensive value of time period differences is greater than the comprehensive value of differences in influencing factors, then increase the fusion weight ratio of time period features and decrease the fusion weight ratio of influencing factor features, so that the sum of the fusion weight ratio of time period features and the fusion weight ratio of influencing factor features is 1. If the comprehensive value of differences in influencing factors is greater than the comprehensive value of time period differences, then increase the fusion weight ratio of influencing factor features and decrease the fusion weight ratio of time period features, so that the sum of the fusion weight ratio of time period features and the fusion weight ratio of influencing factor features is also 1.

[0169] The fusion weights are adjusted based on the relationship between the combined value of time-period differences and the combined value of differences in influencing factors. A larger combined value indicates more significant differences in time-period features across different scenarios, potentially having a greater impact on prediction results. Therefore, the fusion weight of time-period features should be increased, while the fusion weight of influencing factor features should be decreased accordingly, but the sum of the two should be kept constant at 1. Conversely, if the combined value of differences in influencing factors is larger, the fusion weight of influencing factor features should be increased.

[0170] S267, multiply each feature vector of the time period feature by the fusion weight ratio of the time period feature to obtain the weighted time period feature vector; multiply each feature vector of the influencing factor feature by the fusion weight ratio of the influencing factor feature to obtain the weighted influencing factor feature vector.

[0171] Each feature vector in the time-period features is multiplied by its corresponding time-period feature fusion weight to obtain a weighted time-period feature vector. Similarly, the feature vectors of influencing factors are weighted. The weighted feature vector better reflects its importance in the fusion process.

[0172] S268, the weighted time period feature vector and the weighted influencing factor feature vector are superimposed according to the feature dimension to generate a cross-scenario fused load feature vector.

[0173] The weighted time-period feature vector and the weighted influencing factor feature vector are superimposed on the feature dimension. For example, if the dimension of the time-period feature vector is m and the dimension of the influencing factor feature vector is n, and m = n, then the elements at corresponding positions of the two vectors are added to obtain a cross-scene fused load feature vector with dimension m (or n); if m ≠ n, they can be transformed to the same dimension through feature mapping or other methods before being superimposed.

[0174] S269, integrate the cross-scenario fusion load feature vectors corresponding to all time period identifiers and electricity consumption influencing factor types to form cross-scenario fusion load features. Each vector in the cross-scenario fusion load features contains a weighted fusion vector of time period information and influencing factor information.

[0175] All the cross-scenario integrated load feature vectors obtained through the above processing are combined to form the final cross-scenario integrated load feature. This cross-scenario integrated load feature integrates time period information and influencing factor information, and adjusts the weights according to scenario differences, so as to more accurately reflect the characteristics of power load under cross-scenario conditions.

[0176] Continuing with step S200:

[0177] S270, input the cross-scene fused load features into the prediction output layer, verify the load prediction error of the prediction output layer based on the cross-scene fused load features, if the error exceeds the preset range, return to the time period feature extraction layer and the influencing factor feature extraction layer, readjust the feature extraction weights of the corresponding dimensions until the prediction error meets the preset range.

[0178] The cross-scenario fused load features are input into the prediction output layer. The prediction output layer performs load prediction based on the fused features and outputs the prediction result. The prediction result is compared with the actual load data, and the prediction error (such as mean squared error, mean absolute error, etc.) is calculated. If the prediction error exceeds a preset range (such as the mean squared error being greater than a certain threshold), it indicates that the weight adjustment of the current time period feature extraction layer and the influencing factor feature extraction layer is not reasonable enough. It is necessary to return to these two feature extraction layers, re-analyze and adjust the feature extraction weights of the corresponding dimensions. The above weight adjustment, feature fusion and error verification process is repeated until the prediction error meets the preset range.

[0179] S280, when the prediction error meets the preset range, save the adjusted weights of the time period feature extraction layer, the weights of the influencing factor feature extraction layer, and the fusion ratio of the feature fusion layer to obtain the large power load prediction model after weight adjustment.

[0180] Once the prediction error is within the preset range, it indicates that the model's weight adjustment has achieved a good effect. At this point, the adjusted parameters, such as the weights of the time-period feature extraction layer, the weights of the influencing factor feature extraction layer, and the fusion ratio of the feature fusion layer, are saved. These parameters together constitute the large-scale power load prediction model after weight adjustment. This large-scale power load prediction model can better adapt to the power load prediction needs across different scenarios.

[0181] See also Figure 1 :

[0182] The S300 utilizes a large model with adjusted weight parameters and load forecasting adaptation parameters to perform real-time load characteristic optimization and forecasting calculations for the target scenario, generating power load forecasting results for specific electricity consumption types in the target scenario.

[0183] In an optional embodiment of this application, real-time power load data of a specific type of electricity consumption in the target scenario and corresponding real-time power consumption influencing factor data are obtained. Real-time time period correlation features and real-time influencing factor correlation features are obtained through a large model, and real-time fused load features are further generated. Based on the real-time fused load features, load numerical prediction is performed through the large model to form a power load prediction sequence for a specific type of electricity consumption in the target scenario. The load peak value, load valley value, and load average value for each prediction period are extracted from the power load prediction sequence for a specific type of electricity consumption in the target scenario as prediction feature parameters. The power load prediction sequence for a specific type of electricity consumption in the target scenario is integrated with the prediction feature parameters to generate a power load prediction result for a specific type of electricity consumption in the target scenario.

[0184] In an optional embodiment of this application, a load adjustment instruction for power dispatching of the target scenario is generated based on the power load forecasting results of a specific type of electricity consumption in the target scenario. The content of the load adjustment instruction corresponds to the load time period variation pattern in the power load forecasting results of the specific type of electricity consumption in the target scenario.

[0185] In this embodiment, transfer learning is used to perform power scenario parameter adaptation processing. The power load prediction parameters of the large model corresponding to the power load data set of specific power consumption type in the source scenario are combined with the scenario difference features in the cross-scenario power load association feature set and converted into load prediction adaptation parameters that are adapted to the specific power consumption type in the target scenario.

[0186] In this embodiment, real-time power load data of a specific type of electricity consumption in the target scenario and corresponding real-time power consumption influencing factor data are input into the weighted power load prediction model. Real-time load feature optimization and prediction calculation are performed in combination with the load prediction adaptation parameters to generate power load prediction results for the specific type of electricity consumption in the target scenario.

[0187] In this embodiment of the application, a load adjustment instruction for power dispatching of the target scenario is generated based on the power load forecasting result of the specific power consumption type of the target scenario. The content of the load adjustment instruction corresponds to the load time period change pattern in the power load forecasting result of the specific power consumption type of the target scenario.

[0188] Specifically, based on the electricity load forecast results for residential electricity consumption scenarios, the time-period variation patterns of the load are analyzed, such as which periods are peak load and which are off-peak load. Corresponding load regulation instructions are generated based on these patterns. For example, during predicted peak load periods (such as 7 PM to 9 PM), instructions can be issued to increase power supply and activate peak-shaving power plants; during off-peak load periods (such as 1 AM to 4 AM), instructions can be issued to reduce unnecessary generation capacity and implement energy storage. The content of the load regulation instructions closely corresponds to the time-period variation patterns of the load to achieve stable and economical operation of the power system.

[0189] During the data collection process, sensitive data such as residential electricity consumption data are involved. To protect user privacy and prevent data leakage, the following technical measures were adopted: The collected raw data was anonymized, removing user-identifying information (such as name and address); encryption technologies were used to protect data transmission and storage, such as using SSL / TLS protocols for data transmission encryption and AES encryption algorithms for stored data encryption; strict data access control was implemented, allowing only authorized personnel to access and process the data; and data was de-identified, converting sensitive specific values ​​into ranges or fuzzy representations to reduce the risk of privacy leakage without affecting data analysis and model training. Through these technical measures, user privacy and security are protected while conducting electricity load forecasting and analysis.

[0190] Furthermore, regarding step S300:

[0191] S310A extracts power load prediction parameters from the pre-trained power load prediction model corresponding to the power load data set of specific power consumption types in the source scenario. The power load prediction parameters of the large model include time period feature extraction weights, influencing factor feature extraction weights, feature fusion coefficients, and prediction calculation parameters.

[0192] After training, the pre-trained large-scale power load prediction model corresponding to the power load dataset of specific power consumption types in the source scenario will save a series of model parameters. These parameters include the connection weights of each neuron in the time period feature extraction layer (time period feature extraction weights), the connection weights of each neuron in the influencing factor feature extraction layer (influencing factor feature extraction weights), the coefficients used to fuse time period features and influencing factor features in the feature fusion layer (feature fusion coefficients), and the parameters used in the prediction calculation in the prediction output layer (such as linear regression coefficients, bias terms, and prediction calculation parameters). These large-scale model power load prediction parameters can be extracted by reading the model's parameter file.

[0193] S320A extracts scenario difference features from the cross-scenario power load association feature set. The scenario difference features include the difference in load peak value between the source scenario and the target scenario under the same time period, the difference in load change rate under the same meteorological data, and the difference in load fluctuation cycle corresponding to a specific electricity consumption type.

[0194] The cross-scenario power load correlation feature set contains information reflecting the differences between the source and target scenarios, from which scenario difference features are extracted. Peak load difference refers to the difference between the peak load of the source scenario and the peak load of the target scenario under the same time period (i.e., the same time period). Load change rate difference refers to the difference between the rate at which the load of the source scenario changes with meteorological data and the rate at which the load of the target scenario changes with meteorological data under the same meteorological data (e.g., the same temperature range). Load fluctuation cycle difference refers to the difference between the load fluctuation cycle of a specific electricity consumption type (industrial electricity) in the source scenario and the load fluctuation cycle of a specific electricity consumption type (residential electricity) in the target scenario. For example, industrial electricity may have a relatively stable fluctuation cycle on a daily basis, while residential electricity may have different fluctuation cycles on weekends and weekdays.

[0195] S330A performs scenario difference adaptation processing on the time period feature extraction weights in the power load prediction parameters of the large model. Based on the load peak difference in the scenario difference features, it calculates the adjustment ratio of the time period feature extraction weights. If the load peak of a certain time period in the target scenario is higher than the corresponding time period in the source scenario, the time period feature extraction weights corresponding to the corresponding time period are increased; otherwise, the time period feature extraction weights corresponding to the corresponding time period are decreased, thus obtaining the adapted time period feature extraction weights.

[0196] For each time period, an adjustment ratio for the time period feature extraction weights is calculated based on the difference in peak load. For example, the ratio of the peak load of the target scenario to the peak load of the source scenario is calculated. If this ratio is greater than 1, it indicates that the peak load of the target scenario in that time period is higher than that of the source scenario, and the features of that time period may be more important for load prediction of the target scenario. Therefore, the time period feature extraction weights corresponding to that time period are increased according to a certain adjustment ratio. If the ratio is less than 1, the corresponding weights are decreased. The magnitude of the adjustment ratio can be determined based on the degree of difference in peak load; the greater the difference, the larger the adjustment ratio. After the above adjustments, time period feature extraction weights adapted to the target scenario are obtained.

[0197] S340A performs scenario difference adaptation processing on the feature extraction weights of influencing factors in the power load forecasting parameters of large models. Based on the difference in load change rate in the scenario difference features, it calculates the adjustment coefficient of the feature extraction weights of influencing factors. If the load change rate under a certain meteorological data in the target scenario is higher than the change rate under the corresponding meteorological data in the source scenario, the feature extraction weight of the influencing factor corresponding to that meteorological data is increased; otherwise, the feature extraction weight of the influencing factor corresponding to that meteorological data is decreased, thus obtaining the adapted feature extraction weights of influencing factors.

[0198] For different meteorological data types, adjustment coefficients are calculated for the feature extraction weights of influencing factors based on differences in load change rates. For example, under high-temperature meteorological data, the ratio of the load change rate of the target scenario to that of the source scenario is calculated. If this ratio is greater than 1, it indicates that the target scenario is more sensitive to load changes under high-temperature weather conditions, and the feature extraction weights of influencing factors corresponding to high-temperature meteorological data should be increased; the adjustment coefficient can be set according to the magnitude of the ratio. Conversely, the weights are decreased. Similar adjustments are made to the feature extraction weights of influencing factors corresponding to all meteorological data types to obtain the adapted feature extraction weights of influencing factors.

[0199] S350A performs scenario difference adaptation processing on the feature fusion coefficients in the power load prediction parameters of the large model. Based on the load fluctuation cycle difference in the scenario difference features, it calculates the correction value of the feature fusion coefficient. If the load fluctuation cycle of the target scenario is longer than that of the source scenario, it increases the proportion of time period features in the fusion and corrects the feature fusion coefficient. Conversely, it increases the proportion of influencing factor features in the fusion and corrects the feature fusion coefficient, thus obtaining the adapted feature fusion coefficient.

[0200] The correction value for the feature fusion coefficient is calculated based on the difference in load fluctuation cycles. If the load fluctuation cycle of the target scenario is longer than that of the source scenario (e.g., the fluctuation cycle of residential electricity consumption may be affected by weekly patterns and is longer than the daily fluctuation cycle of industrial electricity consumption), it indicates that the changing trend of time-period features over a longer period is more important for predicting the target scenario. Therefore, it is necessary to increase the proportion of time-period features in the fusion, and this adjustment is achieved by correcting the feature fusion coefficient. If the fluctuation cycle of the target scenario is shorter than that of the source scenario, the proportion of influencing factor features is increased. The correction value can be determined based on the proportion of cycle difference; for example, the greater the cycle difference, the larger the correction value, and the greater the adjustment range of the fusion coefficient. After correction, the adapted feature fusion coefficient is obtained.

[0201] S360A performs scenario difference adaptation processing on the prediction calculation parameters in the large model power load prediction parameters. Based on the comprehensive impact of load peak difference, load change rate difference, and load fluctuation cycle difference in the scenario difference characteristics, it calculates the compensation value of the prediction calculation parameters and superimposes the compensation value onto the source scenario prediction calculation parameters to obtain the adapted prediction calculation parameters.

[0202] The impact of differences in peak load, load change rate, and load fluctuation cycle on the prediction calculation parameters is comprehensively considered. By establishing a mapping relationship between these differences and the prediction calculation parameters (e.g., through multiple regression analysis), compensation values ​​for the prediction calculation parameters are calculated. The magnitude and direction of the compensation values ​​depend on the combined effect of these differences. For example, if the target scenario has higher peak loads, a larger change rate, and a longer cycle, it may be necessary to positively compensate for the bias terms or coefficients in the prediction calculation parameters. The calculated compensation values ​​are then superimposed on the prediction calculation parameters of the source scenario to obtain prediction calculation parameters adapted to the target scenario.

[0203] S370A integrates the adapted time period feature extraction weights, adapted influencing factor feature extraction weights, adapted feature fusion coefficients, and adapted prediction calculation parameters to form load prediction adaptation parameters for specific electricity consumption types in the target scenario.

[0204] The adapted parameters obtained from steps S330A to S360A are integrated to form a complete set of load forecasting adaptation parameters. This set contains all model parameters adjusted for the target scenario (residential electricity consumption scenario), enabling the model to better adapt to the electricity consumption characteristics of the target scenario.

[0205] S380A substitutes the load forecasting adaptation parameters into the historical load data sample of the specific electricity consumption type in the target scenario, performs forecasting verification calculation, and if the deviation between the forecasting result and the historical actual load data exceeds the preset threshold, it returns to recalculate the adjustment ratio, adjustment coefficient, correction value and compensation value of each parameter until the deviation meets the preset threshold, and finally determines the load forecasting adaptation parameters.

[0206] Furthermore, regarding step S380A:

[0207] S381A, extract a portion of historical load data from the power load data set of specific electricity consumption type in the target scenario as a historical load verification sample set for the target scenario. The historical load verification sample set for the target scenario includes load values ​​for each time period on multiple historical dates and corresponding historical electricity consumption influencing factor data.

[0208] From historical load data of residential electricity consumption scenarios, a portion of the data is randomly selected or selected according to certain rules (such as covering different seasons, different months, etc.) as a verification sample set. The above sample includes load values ​​for each time period of multiple historical dates (such as some dates in the past year), as well as corresponding historical meteorological data, residential daily life data, and other electricity consumption influencing factors.

[0209] S382A inputs the historical electricity consumption influencing factor data from the historical load verification sample set of the target scenario into the influencing factor feature extraction layer of the weighted power load prediction big model, and combines the weighted influencing factor feature extraction from the adapted load prediction adaptation parameters to generate historical influencing factor features.

[0210] Historical electricity consumption influencing factor data is input into the influencing factor feature extraction layer, along with adapted influencing factor feature extraction weights. The influencing factor feature extraction layer uses these weights to extract features from the historical electricity consumption influencing factor data, generating historical influencing factor features that reflect the impact of historical electricity consumption influencing factors on the load.

[0211] S383A inputs historical time period information from the target scenario historical load verification sample set into the time period feature extraction layer of the weighted power load prediction big model, and combines the adapted time period feature extraction weights from the load prediction adaptation parameters to generate historical time period features.

[0212] Historical time period information (such as time period identifiers) is input into the time period feature extraction layer. Combined with the adapted time period feature extraction weights, historical time period features are extracted, which reflect the load change characteristics of historical time periods.

[0213] S384A inputs the characteristics of historical influencing factors and historical time periods into the feature fusion layer of the power load forecasting model after weight adjustment, and combines the adapted feature fusion coefficients in the load forecasting adaptation parameters to generate historical fused load characteristics.

[0214] Historical influencing factor features and historical time period features are input into the feature fusion layer. The two features are then fused using the adapted feature fusion coefficient to generate historical fused load features, which integrate information from historical time periods and influencing factors.

[0215] S385A inputs historical fusion load characteristics into the prediction output layer of the weighted power load prediction model, combines the adapted prediction calculation parameters in the load prediction adaptation parameters, performs load prediction calculation, and obtains historical load prediction values.

[0216] The historical load characteristics are input into the prediction output layer, and the load prediction is calculated by combining the adapted prediction calculation parameters to obtain the load prediction value for each historical period.

[0217] S386A extracts the actual load values ​​for each time period corresponding to the historical date from the historical load verification sample set of the target scenario, calculates the difference between the historical load prediction value and the actual load value, and obtains the prediction deviation value for each time period.

[0218] The historical load forecast values ​​obtained from the forecast output layer are compared with the actual load values ​​in the sample set on a time-by-time basis, and the difference between the two is calculated, which is the forecast deviation value for each time period.

[0219] S387A: Calculate the average of the absolute values ​​of the prediction deviations for all time periods to obtain the average prediction deviation; calculate the maximum value of the prediction deviations for all time periods to obtain the maximum prediction deviation.

[0220] The prediction deviation values ​​for all time periods are processed, and the average of their absolute values ​​is calculated to obtain the average prediction deviation, which reflects the overall prediction accuracy. The maximum prediction deviation value is calculated to reflect the extreme error of the prediction results.

[0221] S388A compares the average prediction deviation with a preset average deviation threshold and compares the maximum prediction deviation with a preset maximum deviation threshold.

[0222] The preset average deviation threshold and maximum deviation threshold are set according to the prediction requirements of the target scenario and the actual application scenario. For example, the average deviation threshold can be set as a certain percentage of the actual average load, and the maximum deviation threshold can be set as a certain percentage of the actual maximum load.

[0223] S389A: If the average prediction deviation exceeds the preset average deviation threshold, or the maximum prediction deviation exceeds the preset maximum deviation threshold, then the prediction result deviation is determined to exceed the preset threshold. In this case, the source of the deviation is analyzed: if the deviation mainly comes from a certain time period, then the feature extraction weight adjustment ratio corresponding to that time period is recalculated; if the deviation mainly comes from a certain electricity consumption influencing factor, then the influencing factor feature extraction weight adjustment coefficient corresponding to that electricity consumption influencing factor is recalculated; if the deviation comes from the feature fusion stage, then the feature fusion coefficient correction value is recalculated; if the deviation comes from the prediction calculation stage, then the prediction calculation parameter compensation value is recalculated.

[0224] If the average or maximum prediction deviation exceeds a preset threshold, it is necessary to analyze the main sources of the deviation. By analyzing the prediction deviation values ​​for each time period, it can be determined whether the deviation is particularly large in certain time periods, thus identifying that the deviation mainly originates from time period characteristics. By analyzing the prediction deviation under different electricity consumption influencing factors, it can be determined whether there are problems in the handling of certain influencing factors. By analyzing the relationship between the fused features and the prediction results, it can be determined whether there are deficiencies in the feature fusion process. If there are no obvious problems in the above processes, the prediction calculation parameters may need to be adjusted. Based on the analysis results, return to the corresponding steps to recalculate the adjustment ratio, adjustment coefficient, correction value, or compensation value.

[0225] S3810A, after readjusting the corresponding parameters, execute the above historical load forecast calculation and deviation analysis steps again until the average forecast deviation does not exceed the preset average deviation threshold and the maximum forecast deviation does not exceed the preset maximum deviation threshold.

[0226] Based on the newly calculated adjustment ratio, adjustment coefficient, correction value, or compensation value, readjust the corresponding parameters, and then execute steps S382A to S389A again for prediction calculation and deviation analysis. Repeat this process until both the average prediction deviation and the maximum prediction deviation meet the preset threshold requirements.

[0227] S3811A: When the deviation meets the preset threshold, save the current adapted time period feature extraction weight, adapted influencing factor feature extraction weight, adapted feature fusion coefficient, and adapted prediction calculation parameters, and finally determine the load prediction adaptation parameters.

[0228] Once the deviation meets the requirements, save the adjusted adaptation parameters. These parameters are the final determined load forecasting adaptation parameters, which will be used for subsequent real-time load forecasting of the target scenario.

[0229] Continuing with step S300:

[0230] S310B: Obtain real-time power load data of a specific power consumption type in the target scenario through the power data acquisition terminal of the target scenario. The real-time power load data of the specific power consumption type in the target scenario includes the real-time load value of the current period and the historical load value of several consecutive periods.

[0231] In the target scenario of residential electricity consumption, the power data acquisition terminal (such as a smart meter) collects the residential electricity load value in real time for the current period (such as the current 15 minutes), and also stores the historical load value for several consecutive periods (such as 96 periods in the past 24 hours). The above real-time and historical load data constitute the real-time specific electricity load data of the target scenario.

[0232] S320B: Acquire corresponding real-time electricity consumption influencing factor data through environmental data acquisition equipment of the target scene. The real-time electricity consumption influencing factor data includes current meteorological data and real-time work and rest data related to the current electricity consumption type.

[0233] Environmental data acquisition equipment collects real-time meteorological data (such as temperature, humidity, wind speed, etc.) and, by connecting with property management systems and user electricity behavior analysis systems, obtains real-time activity data related to the current type of residential electricity use, such as the current activity of people in the community and whether it is a holiday. The above data serves as real-time electricity consumption influencing factor data and corresponds to real-time load data.

[0234] S330B: Input the historical load values ​​of several consecutive time periods from the real-time specific power load data of the target scenario into the time period feature extraction layer of the power load prediction big model after weight adjustment. Combine the adapted time period feature extraction weights in the load prediction adaptation parameters to extract real-time time period correlation features. The real-time time period correlation features include the load change trend of several consecutive time periods and the time period feature vector corresponding to the current time period.

[0235] Historical load values ​​from multiple consecutive time periods are input into the time period feature extraction layer, which then loads adapted time period feature extraction weights. The time period feature extraction layer processes these historical load values, analyzes load trends (such as upward, downward, or stable trends), and generates a time period feature vector corresponding to the current time period. This information together constitutes the real-time time period correlation features, reflecting the current and recent time period load characteristics.

[0236] S340B: Input the real-time electricity consumption influencing factor data into the influencing factor feature extraction layer of the power load forecasting large model after weight adjustment, combine the adapted influencing factor feature extraction weight in the load forecasting adaptation parameters, and extract the real-time influencing factor correlation features. The real-time influencing factor correlation features include the load influence vector corresponding to the current meteorological data and the load influence vector corresponding to the real-time work and rest data.

[0237] Real-time electricity consumption influencing factor data is input into the influencing factor feature extraction layer, combined with adapted influencing factor feature extraction weights. The influencing factor feature extraction layer processes the current meteorological data and real-time work and rest data respectively, generating a load influence vector corresponding to the current meteorological data (reflecting the degree of influence of meteorological factors on the current load) and a load influence vector corresponding to the real-time work and rest data (reflecting the degree of influence of work and rest factors on the current load). These two vectors together constitute the real-time influencing factor correlation features.

[0238] S350B: Input the real-time time period correlation features and real-time influencing factor correlation features into the feature fusion layer of the power load forecasting big model after weight adjustment, combine them with the adapted feature fusion coefficients in the load forecasting adaptation parameters, perform real-time feature fusion processing, and generate real-time fused load features. The real-time fused load features include a weighted fusion vector of time period information and influencing factor information.

[0239] Real-time time period correlation features and real-time influencing factor correlation features are input into the feature fusion layer, and they are weighted and fused using adapted feature fusion coefficients. During the fusion process, the weight ratio of time period information and influencing factor information is determined according to the feature fusion coefficients, and the feature vectors of the two are superimposed according to their weights to generate a real-time fused load feature, which integrates real-time time period and influencing factor information.

[0240] S360B: Inputs real-time fused load characteristics into the prediction output layer of the power load prediction big model after weight adjustment. Combined with the adapted prediction calculation parameters in the load prediction adaptation parameters, it first predicts the load value for the next period. Then, based on the changing trend of the predicted load value for the next period and the current real-time load value, it predicts the load values ​​for multiple subsequent periods.

[0241] Furthermore, regarding step S360B:

[0242] S361B: Input the real-time fused load characteristics into the prediction output layer of the weighted power load prediction model. The prediction output layer calls the built-in load prediction calculation logic and, in combination with the adapted prediction calculation parameters in the load prediction adaptation parameters, calculates the basic load prediction value for the next period.

[0243] After receiving the real-time fused load characteristics, the prediction output layer calls the internal prediction calculation logic (such as the forward propagation calculation of the neural network), and combines the adapted prediction calculation parameters (such as weights, biases, etc.) to perform preliminary calculations on the load for the next time period (the next 15 minutes) to obtain the basic load prediction value.

[0244] S362B: Extract the load change amplitude of the current time period and several recent consecutive time periods from the real-time time period correlation features, calculate the average value of the load change amplitude, and obtain the recent average load change amplitude.

[0245] The load change amplitude of the current time period and several recent consecutive time periods (such as the past 4 time periods) is obtained from the real-time time period correlation characteristics. The average value of these change amplitudes is calculated to obtain the average value of recent load change amplitude, which reflects the average level of recent load change.

[0246] S363B: Multiply the load base forecast value for the next period by one and the sum of the recent average load change rate to obtain the corrected load forecast value for the next period.

[0247] To improve forecast accuracy, the baseline load forecast is corrected using the average recent load variation. The baseline load forecast is then multiplied by (1 + the average recent load variation) to obtain the corrected load forecast for the next period.

[0248] S364B: Subtract the corrected load forecast value for the next time period from the current real-time load value to obtain the load change difference between the next time period and the current time period.

[0249] The difference between the revised load forecast for the next time period and the current real-time load is calculated, and this difference reflects the change in load in the next time period relative to the current time period.

[0250] S365B: Based on the sign of the load change difference, determine the load change trend for the next period: if the difference is positive, it is determined to be an upward trend; if the difference is negative, it is determined to be a downward trend; if the difference is zero, it is determined to be a stable trend.

[0251] The load change trend for the next period is determined by the sign of the load change difference. A positive difference indicates that the load will increase, a negative difference indicates that the load will decrease, and a zero difference indicates that the load will remain stable.

[0252] S366B: If an upward trend is determined, the rate of increase is calculated based on the load change difference, where the rate of increase is the load change difference divided by the duration of the next time period; based on the rate of increase, the load value for the next time period is predicted as the corrected load prediction value for the next time period plus the rate of increase multiplied by the duration of the next time period.

[0253] If the load change is trending upwards, divide the load change difference by the duration of the next time period (e.g., 15 minutes) to obtain the rate of increase. Then, predict the load value for the next time period by adding the corrected load forecast value for the next time period to the rate of increase multiplied by the duration of the next time period (again, 15 minutes).

[0254] S367B: If a downward trend is determined, the rate of decline is calculated based on the load change difference, where the rate of decline is the absolute value of the load change difference divided by the duration of the next time period; based on the rate of decline, the load value for the next time period is predicted as the corrected load prediction value for the next time period minus the rate of decline multiplied by the duration of the next time period.

[0255] If the load is trending downwards, divide the absolute value of the load change difference by the duration of the next period to obtain the rate of decrease. The predicted load value for the next period is the corrected load forecast for the next period minus the rate of decrease multiplied by the duration of the next period.

[0256] S368B: If the trend is determined to be stable, the predicted load value for the next period will be consistent with the revised load forecast value for the next period.

[0257] If the trend is stable, the load forecast value for the next period will be set to be the same as the revised load forecast value for the next period.

[0258] S369B: Following the above steps, the load values ​​for multiple subsequent time periods are predicted sequentially. The load prediction for each subsequent time period is calculated based on the predicted load value of the previous time period and the corresponding trend.

[0259] Starting from the next time period, following steps S1464 to S1468, load values ​​for multiple subsequent time periods (e.g., the next 24 time periods) are predicted sequentially. The prediction for each time period is based on the prediction results and trends of the previous time period, forming a continuous prediction sequence.

[0260] S3610B: During the forecasting process, if the real-time electricity consumption influencing factor data changes, the real-time influencing factor correlation features are re-extracted, the real-time integrated load features are updated, and the load base forecast value for the next period is recalculated based on the updated real-time integrated load features, thereby adjusting the load forecast value for subsequent periods.

[0261] During the forecasting process, if real-time electricity consumption influencing factors (such as sudden changes in meteorological data or unusual resident schedules) change, the new real-time electricity consumption influencing factor data needs to be re-input into the influencing factor feature extraction layer to extract new real-time influencing factor correlation features. Then, combined with the real-time time period correlation features, the real-time fused load features are updated. The updated real-time fused load features are used to recalculate the load baseline forecast value for the next time period, and the load forecast values ​​for subsequent time periods are adjusted according to the above forecasting steps to ensure that the forecast results can reflect changes in influencing factors in a timely manner.

[0262] S3611B: Arrange all the predicted load values ​​for the subsequent multiple time periods in chronological order to form a preliminary load forecast sequence.

[0263] The predicted load values ​​for subsequent time periods are arranged in chronological order (e.g., starting from the next time period) to form a preliminary load forecast sequence.

[0264] S3612B: Corrects abnormal values ​​in the initial load forecast sequence, refers to the load change trend in the real-time time period correlation characteristics, adjusts abnormal values ​​to values ​​that conform to the trend, and finally obtains load values ​​for multiple subsequent time periods.

[0265] The preliminary load forecast sequence is examined to identify outliers (such as values ​​that significantly deviate from the overall trend). These outliers are then corrected by referencing the load trend in real-time time-period correlation characteristics (e.g., an overall upward trend, but a significant drop in forecast values ​​for a particular period), adjusting them to reasonable values ​​that align with the trend. After correction, the final load values ​​for subsequent time periods are obtained.

[0266] Continuing with step S300:

[0267] S370B: The predicted load values ​​for multiple subsequent time periods are sequenced and arranged in chronological order to form a power load prediction sequence for a specific electricity consumption type in the target scenario.

[0268] The resulting load values ​​for multiple subsequent time periods are organized and arranged chronologically to form a complete power load forecast sequence for a specific electricity consumption type (residential electricity) in the target scenario. This power load forecast sequence demonstrates the predicted residential electricity load over a future period.

[0269] S380B: Extracts the peak load, valley load, and average load for each forecast period from the power load forecast sequence of a specific power consumption type in the target scenario, and uses them as forecast feature parameters.

[0270] For each forecast period in the forecast sequence, the load peak (highest load value), load trough (lowest load value), and average load value within that period are extracted. These parameters can reflect the basic characteristics of the load in each period.

[0271] S390B: Integrate the power load prediction sequence of the specific power consumption type in the target scenario with the prediction feature parameters to generate the power load prediction result of the specific power consumption type in the target scenario. The power load prediction result of the specific power consumption type in the target scenario includes the load value of each prediction period and the corresponding prediction feature parameters.

[0272] The power load forecast sequence and extracted forecast feature parameters (peak, valley, and average values) are integrated to form the final power load forecast result for the specific power consumption type of the target scenario. This power load forecast result for the specific power consumption type of the target scenario comprehensively reflects the future power load situation of the target scenario.

[0273] Combination Figure 2 As shown, this application embodiment provides a cross-scenario power load forecasting device 800 under a large model, including a processor 801 and a memory 802. Optionally, the device may further include a communication interface 803 and a bus 804. The processor 801, communication interface 803, and memory 802 can communicate with each other via the bus 804. The communication interface 803 can be used for information transmission. The processor 801 can call logical instructions in the memory 802 to execute the cross-scenario power load forecasting method under the large model described in the above embodiment.

[0274] Furthermore, the logic instructions in the aforementioned memory 802 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0275] The memory 802, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as the program instructions / modules corresponding to the methods in the embodiments of this application. The processor 801 executes functional applications and data processing by running the program instructions / modules stored in the memory 802, thereby realizing the cross-scenario power load prediction method under the large model in the above embodiments.

[0276] The memory 802 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 802 may include high-speed random access memory and may also include non-volatile memory.

[0277] This application provides a system comprising: a system body and a cross-scenario power load forecasting device 800 under the aforementioned large-scale model. The cross-scenario power load forecasting device 800 under the system large-scale model is installed on the system body. The installation relationship described herein is not limited to placement within the system, but also includes installation and connection with other components of the system, including but not limited to physical connections, electrical connections, or signal transmission connections. Those skilled in the art will understand that the cross-scenario power load forecasting device 800 under the system large-scale model can be adapted to feasible system bodies to achieve other feasible embodiments.

[0278] This application provides a computer-readable storage medium storing computer-executable instructions configured to execute the cross-scenario power load prediction method under the above-described large model.

[0279] The technical solutions of this application embodiment can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this application embodiment. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code.

[0280] The technical solutions of this application embodiment can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes one or more instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in this application embodiment. The aforementioned storage medium can be a non-transitory storage medium, including: USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk, and other media capable of storing program code.

[0281] The foregoing description and accompanying drawings fully illustrate embodiments of this application to enable those skilled in the art to practice them. Other embodiments may include structural, logical, electrical, procedural, and other changes. The embodiments represent only possible variations. Individual components and functions are optional unless explicitly required, and the order of operation may vary. Parts and features of some embodiments may be included in or replace parts and features of other embodiments. Moreover, the terminology used in this application is for describing embodiments only and is not intended to limit the claims. As used in the description of embodiments and claims, the singular forms “a,” “an,” and “the” are intended to equally include the plural forms unless the context clearly indicates otherwise. Similarly, the term “and / or,” as used herein, means including one or more of the associated listed items and all possible combinations thereof. Additionally, when used in this application, the term "comprise" and its variations "comprises" and / or "comprising" refer to the presence of stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components, and / or groups thereof. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, or apparatus that includes said element. In this document, each embodiment may focus on the differences from other embodiments, and similar or identical parts between embodiments can be referred to mutually. For methods, products, etc., of the embodiments claimed, if they correspond to the method section of the embodiments claimed, then the relevant parts can be referred to the description of the method section.

[0282] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments claimed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the embodiments of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0283] The methods and products (including but not limited to devices and equipment) disclosed in the embodiments herein can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units may be merely a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to implement this embodiment according to actual needs. In addition, the functional units in the embodiments of this application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0284] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than that shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. In the descriptions corresponding to the flowcharts and block diagrams in the accompanying drawings, the operations or steps corresponding to different blocks may also occur in a different order than disclosed in the description; sometimes there is no specific order between different operations or steps. For example, two consecutive operations or steps may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. Each block in a block diagram and / or flowchart, and combinations of blocks in a block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

Claims

1. A method for predicting cross-scenario power load under a large model, characterized in that, include: Based on the specific electricity load data of the source and target scenarios and the corresponding electricity influencing factors data, a cross-scenario electricity load association feature set is generated. Based on the cross-scenario power load association feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning. Using the large model with adjusted weight parameters and the load prediction adaptation parameters, real-time load characteristic optimization and prediction calculations are performed for the target scenario to generate power load prediction results for specific electricity consumption types in the target scenario.

2. The method according to claim 1, characterized in that, Based on the specific electricity load data of the source and target scenarios and the corresponding electricity influencing factors, a cross-scenario electricity load correlation feature set is generated, including: Extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type in the source scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the time period association feature of power load in the source scenario. Extract the load value sequence corresponding to the same time period identifier from the power load data set of specific power consumption type of the target scenario, calculate the change range of load value under the same time period identifier on different dates, and generate the time period association feature of power load in the target scenario; The source scenario power load time period association features and the target scenario power load time period association features are mapped one-to-one according to the time period identifier. The similarity parameter between the source scenario power load time period association features and the target scenario power load time period association features under the corresponding time period identifier is calculated to generate cross-scenario load time period association features.

3. The method according to claim 2, characterized in that, Extract the load value sequence corresponding to the same time period identifier from the power load data set of the specific power consumption type in the source scenario, calculate the change range of the load value under the same time period identifier on different dates, and generate the time period association feature of the power load in the source scenario, including: From the power load data set of the specific power consumption type of the source scenario, all load data with the same time period identifier are selected to form the load value sequence of the source scenario. Based on the load values ​​of two adjacent dates in the load value sequence of the source scenario, the load change amplitude of all adjacent dates is calculated to form a load change amplitude sequence. Based on the load change amplitude sequence, the upward trend feature value, the downward trend feature value, and the stable trend feature value are obtained and integrated to form the time period correlation feature of the source scenario power load.

4. The method according to claim 2, characterized in that, Based on the specific electricity load data of the source and target scenarios and the corresponding electricity influencing factors, a cross-scenario electricity load correlation feature set is generated, including: Extract the correlation between meteorological data and industrial production shift data from the source scenario electricity consumption influencing factor data, calculate the load change rate corresponding to industrial production shifts under different meteorological data, and generate the source scenario electricity consumption influencing factor correlation characteristics. Extract the correlation between meteorological data and residents' daily routine data from the data on electricity consumption in the target scenario, calculate the load change rate corresponding to residents' daily routine under different meteorological data, and generate the correlation features of electricity consumption in the target scenario. The correlation features of power consumption influencing factors in the source scenario and the correlation features of power consumption influencing factors in the target scenario are mapped one-to-one according to meteorological data types. Complementary parameters of the correlation features of power consumption influencing factors in the source scenario and the correlation features of power consumption influencing factors in the target scenario under the corresponding meteorological data type are calculated to generate cross-scenario load influencing factor correlation features.

5. The method according to claim 4, characterized in that, Based on the cross-scenario power load correlation feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including: The cross-scenario load time period correlation features are input into the large model. The large model performs hierarchical analysis on the source scenario time period features, target scenario time period features, and similarity parameters in the cross-scenario load time period correlation features to identify significant time period feature dimensions that affect the accuracy of load prediction. Calculate the contribution ratio of the significant time period feature dimension in the source scene and the target scene, and adjust the feature extraction weight of the corresponding dimension in the large model according to the contribution ratio.

6. The method according to claim 5, characterized in that, Based on the cross-scenario power load correlation feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including: The cross-scenario load influencing factor correlation features are input into the large model. The large model performs hierarchical analysis on the source scenario influencing factor features, target scenario influencing factor features and complementary parameters in the cross-scenario load influencing factor correlation features to identify the significant influencing factor feature dimensions for load prediction accuracy. Calculate the complementarity coefficient of the feature dimension of the significant influencing factor in the source scene and the target scene, and adjust the feature extraction weight of the corresponding dimension in the large model according to the complementarity coefficient.

7. The method according to claim 6, characterized in that, Based on the cross-scenario power load correlation feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including: Based on the time period characteristics and influencing factor characteristics of the source and target scenarios after weighting, and based on the scenario difference parameters in the cross-scenario power load association feature set, the fusion ratio of the time period characteristics and influencing factor characteristics is adjusted to generate cross-scenario fused load characteristics. The cross-scenario fused load features are input into the large model to verify the load prediction error of the cross-scenario fused load features. If the error exceeds the preset range, the feature extraction weights of the corresponding dimensions are readjusted until the prediction error meets the preset range.

8. The method according to claim 7, characterized in that, Based on the adjusted weights of the time-period characteristics and influencing factor characteristics of the source and target scenarios, and according to the scenario difference parameters in the cross-scenario power load correlation feature set, the fusion ratio of the time-period characteristics and influencing factor characteristics is adjusted to generate cross-scenario fused load characteristics, including: The time-period features and influencing factor features of the output are obtained from the large model after weight adjustment, and a cross-scenario fused load feature vector is generated based on the time-period features and influencing factor features; The cross-scenario fusion load feature vectors corresponding to all time period identifiers and electricity consumption influencing factor types are integrated to form cross-scenario fusion load features.

9. The method according to claim 7, characterized in that, Based on the cross-scenario power load correlation feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including: The source scenario power load prediction parameters are extracted from the large model. The time period feature extraction weights in the source scenario power load prediction parameters are subjected to scenario difference adaptation processing. Based on the load peak difference in the scenario difference features, the adjustment ratio of the time period feature extraction weights is calculated. If the load peak of the target scenario in a certain time period is higher than that of the corresponding time period in the source scenario, the time period feature extraction weights corresponding to the corresponding time period are increased; otherwise, the time period feature extraction weights corresponding to the corresponding time period are decreased, thus obtaining the adapted time period feature extraction weights.

10. The method according to claim 9, characterized in that, Based on the cross-scenario power load correlation feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including: The influencing factor feature extraction weights in the power load prediction parameters of the source scenario are subjected to scenario difference adaptation processing. Based on the difference in load change rate in the scenario difference features, the adjustment coefficient of the influencing factor feature extraction weights is calculated. If the load change rate under a certain meteorological data in the target scenario is higher than the change rate under the corresponding meteorological data in the source scenario, the influencing factor feature extraction weights corresponding to that meteorological data are increased, and vice versa. The adapted influencing factor feature extraction weights are obtained.

11. The method according to claim 10, characterized in that, Based on the cross-scenario power load correlation feature set, the weight parameters of the pre-trained power load prediction model are adjusted, and the prediction parameters of the source scenario are converted into load prediction adaptation parameters for the target scenario through transfer learning, including: The feature fusion coefficients in the power load prediction parameters of the source scenario are subjected to scenario difference adaptation processing. Based on the load fluctuation cycle difference in the scenario difference features, the correction value of the feature fusion coefficient is calculated. If the load fluctuation cycle of the target scenario is longer than that of the source scenario, the proportion of the time period feature in the fusion is increased and the feature fusion coefficient is corrected. Conversely, the proportion of the influencing factor feature in the fusion is increased and the feature fusion coefficient is corrected to obtain the adapted feature fusion coefficient. The prediction calculation parameters in the source scenario power load prediction parameters are subjected to scenario difference adaptation processing. Based on the comprehensive impact of load peak difference, load change rate difference, and load fluctuation cycle difference in the scenario difference characteristics, the compensation value of the prediction calculation parameters is calculated. The compensation value is superimposed on the source scenario prediction calculation parameters to obtain the adapted prediction calculation parameters. The adapted time period feature extraction weights, influencing factor feature extraction weights, feature fusion coefficients, and prediction calculation parameters are integrated to form load prediction adaptation parameters adapted to the specific electricity consumption type of the target scenario. The load prediction adaptation parameters are substituted into the historical load data sample of the specific electricity consumption type in the target scenario, and prediction verification calculation is performed. If the deviation between the prediction result and the historical actual load data exceeds the preset threshold, the parameter adjustment ratio, adjustment coefficient, correction value and compensation value are recalculated until the deviation meets the preset threshold.

12. The method according to claim 11, characterized in that, Using the adjusted weight parameters of the large model and the load forecasting adaptation parameters, real-time load characteristic optimization and forecasting calculations are performed for the target scenario to generate power load forecasting results for specific electricity consumption types in the target scenario, including: The system acquires real-time power load data for specific power consumption types in the target scenario and corresponding real-time power consumption influencing factor data. It then uses the large model to obtain real-time time period correlation features and real-time influencing factor correlation features, and further generates real-time fused load features. Based on the real-time fused load characteristics, load numerical prediction is performed through the large model to form a power load prediction sequence for a specific type of electricity consumption in the target scenario. The peak load, valley load, and average load for each prediction period are extracted from the power load prediction sequence for the specific power consumption type of the target scenario and used as prediction feature parameters. The power load prediction sequence for the specific power consumption type of the target scenario is then integrated with the prediction feature parameters to generate the power load prediction result for the specific power consumption type of the target scenario.

13. The method according to claim 11, characterized in that, Using the adjusted weight parameters of the large model and the load forecasting adaptation parameters, real-time load characteristic optimization and forecasting calculations are performed for the target scenario to generate power load forecasting results for specific electricity consumption types in the target scenario, including: Based on the power load forecast results for the specific power consumption type of the target scenario, a load adjustment instruction for power dispatching of the target scenario is generated. The content of the load adjustment instruction corresponds to the load time period variation pattern in the power load forecast results for the specific power consumption type of the target scenario.

14. A cross-scenario power load forecasting device under a large model, comprising a processor and a memory storing program instructions, characterized in that, The processor is configured to execute, when running the program instructions, the cross-scenario power load forecasting method under a large model as described in any one of claims 1 to 13.

15. A system, characterized in that, include: System body; as well as, The cross-scenario power load prediction device under the large model as described in claim 14 is installed on the system body.

16. A computer-readable storage medium storing program instructions, characterized in that, When the program instructions are executed, they cause the computer to perform the cross-scenario power load forecasting method under the large model as described in any one of claims 1 to 13.

Images